Your observation that “In its collapsed representation, a subprocess describes the inputs and outputs of the service …” is not entirely correct given that one thing a BPMN diagram doesn’t show is data-flows.

Data (ie inputs and outputs) available to a sub-process might be inferred by the observer and in fact might be defined in sub-process/task attributes, but they are not explicit in a BPMN diagram.

Adding annotations to flow lines is one way of providing this information on a model, but that is somewhat redundant if data definitions are also recorded in process attributes.